已关闭
[Bug-Report|缺陷反馈]: 28处L2_DFX_PHASE_1宏校验缺失参数或顺序不对 #4930
liuhanwen8创建于 9月3日关闭于 15 天前
游震
9月3日 评论：
9月3日 评论：
#宏参数问题清单
| 是否完成 | 序号 | 问题描述 | 算子 | 处理人 |
|---|---|---|---|---|
| [ ] | 1 | DFX_IN：qInputLayout 重复、kvInputLayout 缺失 |
block_sparse_attention_grad | @east_yang |
| [ ] | 2 | DFX_IN：缺失 softmaxOutLayout |
flash_attention_score (V4) | @yu-xinjie62 |
| [ ] | 3 | DFX_IN：缺失 attenMaskOptional、preTokens、nextTokens、sparseMode |
flash_attention_score (Quant) | @yu-xinjie62 |
| [x ] | 4 | DFX_IN：sinkInOptional 顺序错位，应在 prefixOptional 之前 |
flash_attention_score_grad (V3) | @yu-xinjie62 |
| [ ] | 5 | DFX_IN：缺失 activationMode、padSlotId、runMode、maxQueryLen、residualConnection、blockSize、convModeDFX_OUT：顺序不符 |
fused_causal_conv1d | @xiapengcheng5 |
| [ ] | 6 | DFX_IN：缺失同5 + maxDraftTokens（共8个）DFX_OUT：顺序不符 |
fused_causal_conv1d (V2) | @xiapengcheng5 |
| [ ] | 7 | DFX_IN：缺失同5（共7个） DFX_OUT：顺序不符 |
inplace_fused_causal_conv1d | @xiapengcheng5 |
| [ ] | 8 | DFX_IN：缺失同5 + maxDraftTokens（共8个）DFX_OUT：顺序不符 |
inplace_fused_causal_conv1d (V2) | @xiapengcheng5 |
| [ x] | 9 | DFX_IN：outputCache 位置错位，应在末尾 |
nsa_compress_with_cache | @L_Euler |
| [ ] | 10 | DFX_OUT：顺序不符，应为 stateRef, out |
recurrent_gated_delta_rule | @abaabc |
| [ ] | 11 | DFX_OUT：顺序不符，应为 initialStateRef, attnOut, finalState |
recurrent_kda | @jiang-lirui |
| [ ] | 12 | DFX_IN：缺失 softmaxLseFlag、blockShape |
blitz_sparse_attention | @jiang-lirui |
| [ x] | 13 | DFX_IN：sinkInOptional 顺序错位，应在 prefixOptional 之前 |
flash_attention_score_grad (V3, experimental) | @@yu-xinjie62 |
| [ ] | 14 | DFX_IN：缺失 attenMaskOptional、preTokens、nextTokens、sparseMode顺序不符 |
quant_flash_attention_score_grad (experimental) | @yu-xinjie62 |
| [ ] | 15 | DFX_OUT：顺序不符，应为 stateRef, out |
fused_gdn_decode | @jiang-lirui |
| [ ] | 16 | DFX_IN：使用局部变量 layout 代替函数参数 inputLayout |
minimax_sparse_attention_split_kv | @jiang-lirui |
| [ ] | 17 | DFX_IN/OUT：使用 ffnParams.xxx 代替函数参数名含硬编码值 -1、false |
ffn | @chaotang233 |
| [ ] | 18 | DFX_IN/OUT：使用 ffnParams.xxx 代替函数参数名含硬编码值 -1 |
ffn (V2) | @chaotang233 |
| [ ] | 19 | DFX_IN/OUT：使用 ffnParams.xxx 代替函数参数名含硬编码值 -1 |
ffn (V3) | @chaotang233 |
| [ x] | 20 | DFX_IN：缺失 offsetOptional、antiquantScaleOptional、antiquantOffsetOptional |
grouped_matmul_finalize_routing (V2) | @wenxin_fight |
| [ ] | 21 | DFX_IN：缺失 blockIdxFirstScheduledTokenOptional、blockIdxLastScheduledTokenOptional、initialStateIdxOptional、numComputedTokensOptional、activation、nullBlockId、blockSizeToAlign |
causal_conv1d (Fn) | @wangrui_ |
| [ ] | 22 | DFX_IN：缺失 blockIdxLastScheduledTokenOptional、initialStateIdxOptional、activation、nullBlockId、maxQueryLen |
causal_conv1d (Update) | @wangrui_ |
| [ ] | 23 | DFX_IN/OUT：使用 params.xxx 代替函数参数名 |
mhc_post_backward | @xuejinghui |
| [ x] | 24 | DFX_IN：缺失 hcEps |
mhc_pre_backward | @xuejinghui |
| [ x] | 25 | DFX_IN：缺失 hcEps |
mhc_pre_sinkhorn_backward | @xuejinghui |
| [ x] | 26 | DFX_IN：使用 layoutQkvAttr/layoutQOutAttr（局部变量）代替函数参数 layoutQkv/layoutQOutDFX_OUT：顺序不符 |
qkv_rms_norm_rope_cache_with_k_scale | @luwei_1104 |
| [ x] | 27 | DFX_IN：引用不存在的函数参数 cacheMode |
rope_with_sin_cos_cache | @luwei_1104 |
| [ ] | 28 | DFX_OUT：顺序不符，qOut 应在 kCacheRef/vCacheRef 之后 |
und_gen_qkv_rms_norm_rope_cache | @luwei_1104 |
游震
9月3日 评论：
9月3日 评论：
/assign @east_yang 关注下问题1
/assign @yu-xinjie62 关注下问题2、 3、 4、 13、 14、
/assign @xiapengcheng5 关注下问题5、6、7、8
/assign @L_Euler 关注下问题9
/assign @abaabc 关注下问题10
/assign @jiang-lirui 关注下问题12、15 experimental算子且无同名发布算子，找不到处理人
/assign @monologue815 关注下问题11
/assign @renruhai 关注下问题16
/assign @chaotang233 关注下问题17、18、19
/assign @wenxin_fight 关注下问题20
/assign @wangrui_ 关注下问题21、22
/assign @xuejinghui 关注下问题23、24、25
/assign @luwei_1104 关注下问题26、27、28
请完成后在第一列填 [ x]
9月3日 将 east_yang 设为负责人
此处折叠了51条消息 查看更多
16 天前 关联了pull request：Revert "fix(recurrent_kda): 修正 DFX_OUT 参数顺序与接口声明一致"
wang-minbo
15 天前 评论：
15 天前 评论：
好的，当前所有问题已确认或修改完毕，现计划关闭此issue，如果您有什么新的问题或者疑问，欢迎您继续提issue
15 天前 issue状态由 进行中 改变为 已完成
15 天前 关闭了 issue
15 天前 添加了label：wait-feedback
Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
Describe the current behavior / 问题描述 (Mandatory / 必填)
• 宏 APIName：aclnnBlockSparseAttentionGrad
• 接口算子参数：dout, query, key, value, attentionOut, softmaxLse, blockSparseMaskOptional, attenMaskOptional, blockShapeOptional, actualSeqLengthsOptional, actualSeqLengthsKvOptional, qInputLayout, kvInputLayout, numKeyValueHeads, maskType, scaleValue, preTokens, nextTokens, dq, dk, dv
• DFX_IN：dout, query, key, value, attentionOut, softmaxLse, blockSparseMaskOptional, attenMaskOptional, blockShapeOptional, actualSeqLengthsOptional, actualSeqLengthsKvOptional, qInputLayout, qInputLayout, numKeyValueHeads, maskType, scaleValue, preTokens, nextTokens
• DFX_OUT：dq, dk, dv
• ⚠ 问题：缺失: kvInputLayout；重复: qInputLayout
• 宏 APIName：aclnnFlashAttentionScoreV4
• 接口算子参数：query, key, value, realShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, queryRopeOptional, keyRopeOptional, dScaleQOptional, dScaleKOptional, dScaleVOptional, sinkOptional, prefixOptional, actualSeqQLenOptional, actualSeqKvLenOptional, qStartIdxOptional, kvStartIdxOptional, scaleValue, keepProb, preTokens, nextTokens, headNum, inputLayout, innerPrecise, sparseMode, outDtype, pseType, softmaxOutLayout, seed, offset, softmaxMaxOut, softmaxSumOut, softmaxOutOut, attentionOutOut
• DFX_IN：query, key, value, realShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, queryRopeOptional, keyRopeOptional, dScaleQOptional, dScaleKOptional, dScaleVOptional, sinkOptional, prefixOptional, actualSeqQLenOptional, actualSeqKvLenOptional, qStartIdxOptional, kvStartIdxOptional, scaleValue, keepProb, preTokens, nextTokens, headNum, inputLayout, innerPrecise, sparseMode, outDtype, pseType, seed, offset
• DFX_OUT：softmaxMaxOut, softmaxSumOut, softmaxOutOut, attentionOutOut
• ⚠ 问题：缺失: softmaxOutLayout
• 宏 APIName：aclnnQuantFlashAttentionScore
• 接口算子参数：query, key, value, attenMaskOptional, dScaleQ, dScaleK, dScaleV, pScale, scaleValue, preTokens, nextTokens, headNum, inputLayout, sparseMode, softmaxMaxOut, softmaxSumOut, softmaxOutout, attentionOutOut
• DFX_IN：query, key, value, dScaleQ, dScaleK, dScaleV, pScale, scaleValue, headNum, inputLayout
• DFX_OUT：softmaxMaxOut, softmaxSumOut, softmaxOutout, attentionOutOut
• ⚠ 问题：缺失: attenMaskOptional, preTokens, nextTokens, sparseMode
• 宏 APIName：aclnnFlashAttentionScoreGradV3
• 接口算子参数：query, keyIn, value, dy, pseShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, attentionInOptional, sinkInOptional, prefixOptional, qStartIdxOptional, kvStartIdxOptional, scaleValue, keepProb, preTokens, nextTokens, headNum, inputLayout, innerPrecise, sparseMode, pseType, dqOut, dkOut, dvOut, dpseOut, dsinkOut
• DFX_IN：query, keyIn, value, dy, pseShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, attentionInOptional, prefixOptional, qStartIdxOptional, kvStartIdxOptional, scaleValue, keepProb, preTokens, nextTokens, headNum, inputLayout, innerPrecise, sparseMode, pseType, sinkInOptional
• DFX_OUT：dqOut, dkOut, dvOut, dpseOut, dsinkOut
• ⚠ 问题：IN 顺序不符
• 宏 APIName：aclnnFusedCausalConv1d
• 接口算子参数：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx, activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode, y
• DFX_IN：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx
• DFX_OUT：y, convStates
• ⚠ 问题：缺失: activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode；OUT 顺序不符
• 宏 APIName：aclnnFusedCausalConv1dV2
• 接口算子参数：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx, activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode, maxDraftTokens, y
• DFX_IN：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx
• DFX_OUT：y, convStates
• ⚠ 问题：缺失: activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode, maxDraftTokens；OUT 顺序不符
• 宏 APIName：aclnnInplaceFusedCausalConv1d
• 接口算子参数：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx, activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode
• DFX_IN：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx
• DFX_OUT：convStates, x
• ⚠ 问题：缺失: activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode；OUT 顺序不符
• 宏 APIName：aclnnInplaceFusedCausalConv1dV2
• 接口算子参数：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx, activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode, maxDraftTokens
• DFX_IN：x, weight, convStates, queryStartLoc, cacheIndices, initialStateMode, bias, numAcceptedTokens, numComputedTokens, blockIdxFirstScheduledToken, blockIdxLastScheduledToken, initialStateIdx
• DFX_OUT：convStates, x
• ⚠ 问题：缺失: activationMode, padSlotId, runMode, maxQueryLen, residualConnection, blockSize, convMode, maxDraftTokens；OUT 顺序不符
• 宏 APIName：aclnnNsaCompressWithCache
• 接口算子参数：input, weight, slotMapping, actSeqLenOptional, blockTableOptional, layoutOptional, compressBlockSize, compressStride, actSeqLenType, pageBlockSize, outputCache
• DFX_IN：input, weight, slotMapping, outputCache, actSeqLenOptional, blockTableOptional, layoutOptional, compressBlockSize, compressStride, actSeqLenType, pageBlockSize
• DFX_OUT：outputCache
• ⚠ 问题：IN 顺序不符
• 宏 APIName：aclnnRecurrentGatedDeltaRule
• 接口算子参数：query, key, value, beta, stateRef, actualSeqLengths, ssmStateIndices, g, gk, numAcceptedTokens, scaleValue, out
• DFX_IN：query, key, value, beta, stateRef, actualSeqLengths, ssmStateIndices, g, gk, numAcceptedTokens, scaleValue
• DFX_OUT：out, stateRef
• ⚠ 问题：OUT 顺序不符
• 宏 APIName：aclnnRecurrentKda
• 接口算子参数：query, key, value, gate, beta, initialStateRef, cuSeqlensOptional, ssmStateIndicesOptional, aLogOptional, dtBiasOptional, numAcceptedTokensOptional, layout, scale, outputFinalState, inplaceFinalState, useQkL2normInKernel, useGateInKernel, useBetaSigmoidInKernel, allowNegEigval, safeGate, lowerBound, stateVFirst, attnOut, finalState
• DFX_IN：query, key, value, gate, beta, initialStateRef, cuSeqlensOptional, ssmStateIndicesOptional, aLogOptional, dtBiasOptional, numAcceptedTokensOptional, layout, scale, outputFinalState, inplaceFinalState, useQkL2normInKernel, useGateInKernel, useBetaSigmoidInKernel, allowNegEigval, safeGate, lowerBound, stateVFirst
• DFX_OUT：attnOut, initialStateRef, finalState
• ⚠ 问题：OUT 顺序不符
• 宏 APIName：InnerBlitzSparseAttention
• 接口算子参数：query, key, value, pseShift, attenMask, sabi, actualSeqLengths, actualSeqLengthsKv, deqScale1, quantScale1, deqScale2, quantScale2, quantOffset2, numHeads, scaleValue, preTokens, nextTokens, inputLayout, numKeyValueHeads, sparseMode, innerPrecise, softmaxLseFlag, blockShape, attentionOut, softmaxLse
• DFX_IN：query, key, value, pseShift, attenMask, sabi, actualSeqLengths, actualSeqLengthsKv, deqScale1, quantScale1, deqScale2, quantScale2, quantOffset2, numHeads, scaleValue, preTokens, nextTokens, inputLayout, numKeyValueHeads, sparseMode, innerPrecise
• DFX_OUT：attentionOut, softmaxLse
• ⚠ 问题：缺失: softmaxLseFlag, blockShape
• 宏 APIName：aclnnFlashAttentionScoreGradV3
• 接口算子参数：query, keyIn, value, dy, pseShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, attentionInOptional, sinkInOptional, prefixOptional, qStartIdxOptional, kvStartIdxOptional, scaleValue, keepProb, preTokens, nextTokens, headNum, inputLayout, innerPrecise, sparseMode, pseType, dqOut, dkOut, dvOut, dpseOut, dsinkOut
• DFX_IN：query, keyIn, value, dy, pseShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, attentionInOptional, prefixOptional, qStartIdxOptional, kvStartIdxOptional, scaleValue, keepProb, preTokens, nextTokens, headNum, inputLayout, innerPrecise, sparseMode, pseType, sinkInOptional
• DFX_OUT：dqOut, dkOut, dvOut, dpseOut, dsinkOut
• ⚠ 问题：IN 顺序不符
• 宏 APIName：aclnnQuantFlashAttentionScoreGrad
• 接口算子参数：query, keyIn, value, dy, attenMaskOptional, softmaxMax, softmaxSum, attentionIn, dScaleQ, dScaleK, dScaleV, dScaleDy, dsScale, pScale, scaleValue, preTokens, nextTokens, headNum, inputLayout, sparseMode, outDtype, dqOut, dkOut, dvOut
• DFX_IN：query, keyIn, value, dy, softmaxMax, softmaxSum, attentionIn, dScaleQ, dScaleK, dScaleV, dScaleDy, scaleValue, headNum, inputLayout, outDtype, dsScale, pScale
• DFX_OUT：dqOut, dkOut, dvOut
• ⚠ 问题：缺失: attenMaskOptional, preTokens, nextTokens, sparseMode；IN 顺序不符
• 宏 APIName：aclnnFusedGdnDecode
• 接口算子参数：mixedQkv, a, b, aLog, dtBias, stateRef, ssmStateIndices, scale, softplusThreshold, out
• DFX_IN：mixedQkv, a, b, aLog, dtBias, stateRef, ssmStateIndices, scale, softplusThreshold
• DFX_OUT：out, stateRef
• ⚠ 问题：OUT 顺序不符
• 宏 APIName：aclnnMinimaxSparseAttentionSplitKv
• 接口算子参数：query, key, value, blockTable, k2qRowPtr, k2qQIndices, k2qSlotIndices, actualSeqLengths, actualSeqLengthsKv, numKeyValueHeads, scaleValue, blockSize, topK, innerPrecise, softmaxLseFlag, inputLayout, attentionOut, softmaxLse
• DFX_IN：query, key, value, blockTable, k2qRowPtr, k2qQIndices, k2qSlotIndices, actualSeqLengths, actualSeqLengthsKv, numKeyValueHeads, scaleValue, blockSize, topK, innerPrecise, softmaxLseFlag, layout
• DFX_OUT：attentionOut, softmaxLse
• ⚠ 问题：缺失: inputLayout；多余: layout
• 宏 APIName：aclnnFFN
• 接口算子参数：x, weight1, weight2, expertTokens, bias1, bias2, scale, offset, deqScale1, deqScale2, antiquantScale1, antiquantScale2, antiquantOffset1, antiquantOffset2, activation, innerPrecise, y
• DFX_IN：ffnParams.x, ffnParams.weight1, ffnParams.weight2, expertTokens, ffnParams.bias1, ffnParams.bias2, ffnParams.scale, ffnParams.offset, ffnParams.deqScale1, ffnParams.deqScale2, ffnParams.antiquantScale1, ffnParams.antiquantScale2, ffnParams.antiquantOffset1, ffnParams.antiquantOffset2, activation, ffnParams.innerPrecise, -1, false
• DFX_OUT：ffnParams.y
• ⚠ 问题：缺失: x, weight1, weight2, bias1, bias2, scale, offset, deqScale1, deqScale2, antiquantScale1, antiquantScale2, antiquantOffset1, antiquantOffset2, innerPrecise, y；多余: ffnParams.x, ffnParams.weight1, ffnParams.weight2, ffnParams.bias1, ffnParams.bias2, ffnParams.scale, ffnParams.offset, ffnParams.deqScale1, ffnParams.deqScale2, ffnParams.antiquantScale1, ffnParams.antiquantScale2, ffnParams.antiquantOffset1, ffnParams.antiquantOffset2, ffnParams.innerPrecise, -1, false, ffnParams.y
• 宏 APIName：aclnnFFNV2
• 接口算子参数：x, weight1, weight2, expertTokens, bias1, bias2, scale, offset, deqScale1, deqScale2, antiquantScale1, antiquantScale2, antiquantOffset1, antiquantOffset2, activation, innerPrecise, tokensIndexFlag, y
• DFX_IN：ffnParams.x, ffnParams.weight1, ffnParams.weight2, expertTokens, ffnParams.bias1, ffnParams.bias2, ffnParams.scale, ffnParams.offset, ffnParams.deqScale1, ffnParams.deqScale2, ffnParams.antiquantScale1, ffnParams.antiquantScale2, ffnParams.antiquantOffset1, ffnParams.antiquantOffset2, activation, ffnParams.innerPrecise, -1, tokensIndexFlag
• DFX_OUT：ffnParams.y
• ⚠ 问题：缺失: x, weight1, weight2, bias1, bias2, scale, offset, deqScale1, deqScale2, antiquantScale1, antiquantScale2, antiquantOffset1, antiquantOffset2, innerPrecise, y；多余: ffnParams.x, ffnParams.weight1, ffnParams.weight2, ffnParams.bias1, ffnParams.bias2, ffnParams.scale, ffnParams.offset, ffnParams.deqScale1, ffnParams.deqScale2, ffnParams.antiquantScale1, ffnParams.antiquantScale2, ffnParams.antiquantOffset1, ffnParams.antiquantOffset2, ffnParams.innerPrecise, -1, ffnParams.y
• 宏 APIName：aclnnFFNV3
• 接口算子参数：x, weight1, weight2, expertTokensOptional, bias1Optional, bias2Optional, scaleOptional, offsetOptional, deqScale1Optional, deqScale2Optional, antiquantScale1Optional, antiquantScale2Optional, antiquantOffset1Optional, antiquantOffset2Optional, activation, innerPrecise, tokensIndexFlag, y
• DFX_IN：ffnParams.x, ffnParams.weight1, ffnParams.weight2, expertTokensOptional, ffnParams.bias1, ffnParams.bias2, ffnParams.scale, ffnParams.offset, ffnParams.deqScale1, ffnParams.deqScale2, ffnParams.antiquantScale1, ffnParams.antiquantScale2, ffnParams.antiquantOffset1, ffnParams.antiquantOffset2, activation, ffnParams.innerPrecise, -1, tokensIndexFlag
• DFX_OUT：ffnParams.y
• ⚠ 问题：缺失: x, weight1, weight2, bias1Optional, bias2Optional, scaleOptional, offsetOptional, deqScale1Optional, deqScale2Optional, antiquantScale1Optional, antiquantScale2Optional, antiquantOffset1Optional, antiquantOffset2Optional, innerPrecise, y；多余: ffnParams.x, ffnParams.weight1, ffnParams.weight2, ffnParams.bias1, ffnParams.bias2, ffnParams.scale, ffnParams.offset, ffnParams.deqScale1, ffnParams.deqScale2, ffnParams.antiquantScale1, ffnParams.antiquantScale2, ffnParams.antiquantOffset1, ffnParams.antiquantOffset2, ffnParams.innerPrecise, -1, ffnParams.y
20. [repos/ops-transformer/gmm/grouped_matmul_finalize_routing/op_api/aclnn_grouped_matmul_finalize_routing.cpp:1367] aclnnGroupedMatmulFinalizeRoutingV2GetWorkspaceSize• 宏 APIName：aclnnGroupedMatmulFinalizeRoutingV2
• 接口算子参数：x1, x2, scaleOptional, biasOptional, offsetOptional, antiquantScaleOptional, antiquantOffsetOptional, pertokenScaleOptional, groupListOptional, sharedInputOptional, logitOptional, rowIndexOptional, dtype, sharedInputWeight, sharedInputOffset, transposeX1, transposeX2, groupListType, out
• DFX_IN：x1, x2, scaleOptional, biasOptional, pertokenScaleOptional, groupListOptional, sharedInputOptional, logitOptional, rowIndexOptional, dtype, sharedInputWeight, sharedInputOffset, transposeX1, transposeX2, groupListType
• DFX_OUT：out
• ⚠ 问题：缺失: offsetOptional, antiquantScaleOptional, antiquantOffsetOptional
• 宏 APIName：aclnnCausalConv1dFn
• 接口算子参数：x, weight, convStatesRef, biasOptional, queryStartLocOptional, cacheIndicesOptional, initialStateModeOptional, blockIdxFirstScheduledTokenOptional, blockIdxLastScheduledTokenOptional, initialStateIdxOptional, numComputedTokensOptional, activation, nullBlockId, blockSizeToAlign, y
• DFX_IN：x, weight, convStatesRef, biasOptional, queryStartLocOptional, cacheIndicesOptional, initialStateModeOptional
• DFX_OUT：convStatesRef, y
• ⚠ 问题：缺失: blockIdxFirstScheduledTokenOptional, blockIdxLastScheduledTokenOptional, initialStateIdxOptional, numComputedTokensOptional, activation, nullBlockId, blockSizeToAlign
• 宏 APIName：aclnnCausalConv1dUpdate
• 接口算子参数：x, weight, convStatesRef, biasOptional, queryStartLocOptional, cacheIndicesOptional, numAcceptedTokensOptional, blockIdxLastScheduledTokenOptional, initialStateIdxOptional, activation, nullBlockId, maxQueryLen, y
• DFX_IN：x, weight, convStatesRef, biasOptional, queryStartLocOptional, cacheIndicesOptional, numAcceptedTokensOptional
• DFX_OUT：convStatesRef, y
• ⚠ 问题：缺失: blockIdxLastScheduledTokenOptional, initialStateIdxOptional, activation, nullBlockId, maxQueryLen
• 宏 APIName：aclnnMhcPostBackward
• 接口算子参数：gradOutput, x, hRes, hOut, hPost, gradX, gradHres, gradHout, gradHpost
• DFX_IN：params.gradOutput, params.x, params.hRes, params.hOut, params.hPost
• DFX_OUT：params.gradX, params.gradHres, params.gradHout, params.gradHpost
• ⚠ 问题：缺失: gradOutput, x, hRes, hOut, hPost, gradX, gradHres, gradHout, gradHpost；多余: params.gradOutput, params.x, params.hRes, params.hOut, params.hPost, params.gradX, params.gradHres, params.gradHout, params.gradHpost
• 宏 APIName：aclnnMhcPreBackward
• 接口算子参数：x, phi, alpha, gradHIn, gradHPost, gradHRes, invRms, hMix, hPre, hPost, gamma, gradXPostOptional, hcEps, gradX, gradPhi, gradAlpha, gradBias, gradGamma
• DFX_IN：x, phi, alpha, gradHIn, gradHPost, gradHRes, invRms, hMix, hPre, hPost, gamma, gradXPostOptional
• DFX_OUT：gradX, gradPhi, gradAlpha, gradBias, gradGamma
• ⚠ 问题：缺失: hcEps
• 宏 APIName：aclnnMhcPreSinkhornBackward
• 接口算子参数：gradHin, gradHPost, gradHRes, x, phi, alpha, bias, hPre, hcBeforeNorm, invRms, sumOut, normOut, hcEps, gradX, gradPhi, gradAlpha, gradBias
• DFX_IN：gradHin, gradHPost, gradHRes, x, phi, alpha, bias, hPre, hcBeforeNorm, invRms, sumOut, normOut
• DFX_OUT：gradX, gradPhi, gradAlpha, gradBias
• ⚠ 问题：缺失: hcEps
26. [repos/ops-transformer/posembedding/qkv_rms_norm_rope_cache_with_k_scale/op_host/op_api/aclnn_qkv_rms_norm_rope_cache_with_k_scale.cpp:269] aclnnQkvRmsNormRopeCacheWithKScaleGetWorkspaceSize• 宏 APIName：aclnnQkvRmsNormRopeCacheWithKScale
• 接口算子参数：qkv, qGamma, kGamma, cosSin, slotMapping, kCacheRef, vCacheRef, kScaleCacheRef, queryStartLocOptional, seqLensOptional, rotationOptional, vScaleOptional, mropePositionOptional, headNums, layoutQkv, layoutQOut, epsilon, mropeSectionOptional, qQuantMode, kQuantMode, qOut, qScaleOptional
• DFX_IN：qkv, qGamma, kGamma, cosSin, slotMapping, kCacheRef, vCacheRef, kScaleCacheRef, queryStartLocOptional, seqLensOptional, rotationOptional, vScaleOptional, mropePositionOptional, headNums, layoutQkvAttr, layoutQOutAttr, epsilon, mropeSectionOptional, qQuantMode, kQuantMode
• DFX_OUT：qOut, qScaleOptional, kCacheRef, vCacheRef, kScaleCacheRef
• ⚠ 问题：缺失: layoutQkv, layoutQOut；多余: layoutQkvAttr, layoutQOutAttr；OUT 顺序不符
27. [repos/ops-transformer/posembedding/rope_with_sin_cos_cache/op_host/op_api/aclnn_rope_with_sin_cos_cache.cpp:275] aclnnRopeWithSinCosCacheGetWorkspaceSize• 宏 APIName：aclnnRopeWithSinCosCache
• 接口算子参数：positions, queryIn, keyIn, cosSinCache, mropeSection, headSize, isNeoxStyle, queryOut, keyOut
• DFX_IN：positions, queryIn, keyIn, cosSinCache, mropeSection, headSize, isNeoxStyle, cacheMode
• DFX_OUT：queryOut, keyOut
• ⚠ 问题：多余: cacheMode
• 宏 APIName：aclnnUndGenQkvRmsNormRopeCache
• 接口算子参数：undQkv, undWeightsQ, undWeightsK, cosSinCache, kCacheRef, vCacheRef, slotMapping, positions, genQkvOptional, genWeightsQOptional, genWeightsKOptional, catIndicesOptional, numHeadsQ, numHeadsK, numHeadsV, normEps, mropeSection, qOut
• DFX_IN：undQkv, undWeightsQ, undWeightsK, cosSinCache, kCacheRef, vCacheRef, slotMapping, positions, genQkvOptional, genWeightsQOptional, genWeightsKOptional, catIndicesOptional, numHeadsQ, numHeadsK, numHeadsV, normEps, mropeSection
• DFX_OUT：qOut, kCacheRef, vCacheRef
• ⚠ 问题：OUT 顺序不符
Environment / 环境信息 (Mandatory / 必填)
不涉及
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
不涉及
Describe the expected behavior / 预期结果 (Mandatory / 必填)
L2_DFX_PHASE_1宏使用说明：
https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/latest/API/aolapi/atlasascendc_api_07_1052.html
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
不涉及
Special notes for this issue/备注 (Optional / 选填)